vllm.model_executor.layers.quantization.quark.quark_moe ¶

class QuarkNvfp4MoEMethod(QuarkMoEMethod):
    def __init__(
        self,
        weight_config: dict[str, Any],
        input_config: dict[str, Any],
        moe: FusedMoEConfig,
        quant_config: "QuarkConfig",  # type: ignore # noqa E501 # noqa F821
    ):
        super().__init__(moe)
        self.weight_quant = weight_config
        self.input_quant = input_config
        self.quant_config = quant_config
        self.group_size = 16

        # Select experts implementation.
        self.nvfp4_backend, self.experts_cls = select_nvfp4_moe_backend(
            config=self.moe,
            weight_key=kNvfp4Static,
            activation_key=kNvfp4Dynamic,
        )

    def create_weights(
        self,
        layer: torch.nn.Module,
        num_experts: int,
        hidden_size: int,
        intermediate_size_per_partition: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ):
        layer.num_experts = num_experts
        layer.params_dtype = params_dtype
        layer.quant_config = self.quant_config
        weight_dtype = torch.uint8
        weight_scale_dtype = torch.float8_e4m3fn

        # GEMM 1 - w13 weight
        w13_weight = torch.nn.Parameter(
            torch.empty(
                num_experts,
                self.moe.w13_num_shards * intermediate_size_per_partition,
                # 2 fp4 items are packed in the input dimension
                hidden_size // 2,
                dtype=weight_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight", w13_weight)
        set_weight_attrs(w13_weight, extra_weight_attrs)

        # GEMM 2 - w2 weight
        w2_weight = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                # 2 fp4 items are packed in the input dimension
                intermediate_size_per_partition // 2,
                dtype=weight_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_weight", w2_weight)
        set_weight_attrs(w2_weight, extra_weight_attrs)

        # Weight scales (per-group FP8 scales)
        w13_weight_scale = torch.nn.Parameter(
            torch.empty(
                num_experts,
                self.moe.w13_num_shards * intermediate_size_per_partition,
                hidden_size // self.group_size,
                dtype=weight_scale_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight_scale", w13_weight_scale)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.GROUP.value}
        )
        set_weight_attrs(w13_weight_scale, extra_weight_attrs)

        w2_weight_scale = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                intermediate_size_per_partition // self.group_size,
                dtype=weight_scale_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_weight_scale", w2_weight_scale)
        set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        # Global weight scales (per-tensor FP32 scales)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
        )

        w13_weight_scale_2 = torch.nn.Parameter(
            torch.empty(num_experts, self.moe.w13_num_shards, dtype=torch.float32),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight_scale_2", w13_weight_scale_2)
        set_weight_attrs(w13_weight_scale_2, extra_weight_attrs)

        w2_weight_scale_2 = torch.nn.Parameter(
            torch.empty(num_experts, dtype=torch.float32),
            requires_grad=False,
        )
        layer.register_parameter("w2_weight_scale_2", w2_weight_scale_2)
        set_weight_attrs(w2_weight_scale_2, extra_weight_attrs)

        # Input global scales (per-tensor FP32 scales)
        w13_input_scale_2 = torch.nn.Parameter(
            torch.empty(num_experts, self.moe.w13_num_shards, dtype=torch.float32),
            requires_grad=False,
        )
        layer.register_parameter("w13_input_scale_2", w13_input_scale_2)
        set_weight_attrs(w13_input_scale_2, extra_weight_attrs)

        w2_input_scale_2 = torch.nn.Parameter(
            torch.empty(num_experts, dtype=torch.float32),
            requires_grad=False,
        )
        layer.register_parameter("w2_input_scale_2", w2_input_scale_2)
        set_weight_attrs(w2_input_scale_2, extra_weight_attrs)

    def process_weights_after_loading(self, layer: RoutedExperts) -> None:
        """
        Convert NVFP4 MoE weights into kernel format and setup the kernel.
        """

        # Match existing NVFP4 MoE paths: fused w13 uses the w1 global scale.
        if self.moe.is_act_and_mul and not torch.allclose(
            layer.w13_weight_scale_2[:, 0], layer.w13_weight_scale_2[:, 1]
        ):
            logger.warning_once(
                "w1_weight_scale_2 must match w3_weight_scale_2. "
                "Accuracy may be affected."
            )

        w13_weight_scale_2 = layer.w13_weight_scale_2[:, 0].contiguous()

        w2_weight_scale_2 = layer.w2_weight_scale_2

        (
            w13,
            w13_scale,
            w13_scale_2,
            a13_scale,
            w2,
            w2_scale,
            w2_scale_2,
            a2_scale,
        ) = convert_to_nvfp4_moe_kernel_format(
            nvfp4_backend=self.nvfp4_backend,
            layer=layer,
            w13=layer.w13_weight,
            w13_scale=layer.w13_weight_scale,
            w13_scale_2=w13_weight_scale_2,
            a13_scale=layer.w13_input_scale_2,
            w2=layer.w2_weight,
            w2_scale=layer.w2_weight_scale,
            w2_scale_2=w2_weight_scale_2,
            a2_scale=layer.w2_input_scale_2,
            is_act_and_mul=self.moe.is_act_and_mul,
        )

        replace_parameter(layer, "w13_weight", w13)
        replace_parameter(layer, "w13_weight_scale", w13_scale)
        replace_parameter(layer, "w13_weight_scale_2", w13_scale_2)
        replace_parameter(layer, "w13_input_scale_2", a13_scale)

        replace_parameter(layer, "w2_weight", w2)
        replace_parameter(layer, "w2_weight_scale", w2_scale)
        replace_parameter(layer, "w2_weight_scale_2", w2_scale_2)
        replace_parameter(layer, "w2_input_scale_2", a2_scale)

        # Setup modular kernel.
        self.moe_quant_config = self.get_fused_moe_quant_config(layer)
        if self.moe_quant_config:
            assert self.experts_cls is not None
            self.moe_kernel = make_nvfp4_moe_kernel(
                moe_quant_config=self.moe_quant_config,
                moe_config=self.moe,
                experts_cls=self.experts_cls,
                backend=self.nvfp4_backend,
                routing_tables=layer._expert_routing_tables(),
            )

    def get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        return make_nvfp4_moe_quant_config(
            backend=self.nvfp4_backend,
            w13_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            w13_scale_2=layer.w13_weight_scale_2,
            w2_scale_2=layer.w2_weight_scale_2,
            a13_scale=layer.w13_input_scale_2,
            a2_scale=layer.w2_input_scale_2,
            swiglu_limit=getattr(layer, "swiglu_limit", None),
            swiglu_alpha=getattr(layer, "swiglu_alpha", None),
            swiglu_beta=getattr(layer, "swiglu_beta", None),
            layer=layer,
        )

    def apply(
        self,
        layer: RoutedExperts,
        x: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        shared_experts: SharedExperts | None,
        shared_experts_input: torch.Tensor | None,
    ) -> torch.Tensor:
        assert self.moe_kernel is not None
        return self.moe_kernel.apply(
            x,
            layer.w13_weight,
            layer.w2_weight,
            topk_weights,
            topk_ids,
            activation=layer.activation,
            global_num_experts=layer.global_num_experts,
            expert_map=layer.expert_map,
            apply_router_weight_on_input=layer.apply_router_weight_on_input,
            shared_experts=shared_experts,
            shared_experts_input=shared_experts_input,
        )

class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
    def __init__(
        self,
        weight_config: dict[str, Any],
        input_config: dict[str, Any] | None,
        moe: FusedMoEConfig,
    ):
        super().__init__(moe)
        self.weight_quant = weight_config
        self.input_quant = input_config

        weight_qscheme = self.weight_quant.get("qscheme")
        if not weight_qscheme == "per_group":
            raise ValueError(
                "For MX(FP4) Fused MoE layers, only per-group scales "
                f"for weights are supported. Found {weight_qscheme}."
            )  # noqa E501

        self.weight_dtype = self.weight_quant["dtype"].replace("fp", "mxfp")
        if self.input_quant is not None:
            input_quant = self.input_quant["dtype"]
            if input_quant in ["fp4", "fp6_e3m2", "fp6_e2m3"]:
                self.input_dtype = input_quant.replace("fp", "mxfp")
            elif input_quant == "fp8_e4m3":
                self.input_dtype = input_quant.replace("fp8_e4m3", "fp8")
            else:
                raise NotImplementedError(
                    f"Current input dtype {input_quant} is not compatible \
                        with OCP MX (weight) MoE quantization. Please open an issue"
                )
        else:
            self.input_dtype = None

        self.fp4_dtype = getattr(torch, "float4_e2m1fn_x2", None)

        self.ocp_mx_scheme = OCP_MX_Scheme.from_quant_dtype(
            self.input_dtype, self.weight_dtype
        )

        if self.ocp_mx_scheme is None:
            raise ValueError(
                f"Unsupported OCP MX dtype combination for MoE: "
                f"input_dtype={self.input_dtype}, weight_dtype={self.weight_dtype}. "
                f"Please check that the combination is supported in OCP_MX_Scheme."
            )

        # TODO(bowenbao): refactor and introduce backends for other OCP MX schemes,
        # use kernel abstraction for all OCP MX MOE implementations.
        self.mxfp4_backend: Mxfp4MoeBackend = Mxfp4MoeBackend.NONE
        self.experts_cls: type[mk.FusedMoEExperts] | None = None
        self.moe_kernel: mk.FusedMoEKernel | None = None

        # Used for triton kernel precision configs (W4A8, TRITON backends)
        self.w13_precision_config = None
        self.w2_precision_config = None

        if self.input_quant is not None:
            self.static_input_scales = not self.input_quant.get("is_dynamic")
        else:
            self.static_input_scales = False

        # Select backend based on OCP MX scheme
        if self.ocp_mx_scheme == "w_mxfp4":
            # W4A16: weight-only MXFP4
            self.mxfp4_backend, self.experts_cls = select_mxfp4_moe_backend(moe)
        elif self.ocp_mx_scheme == "w_mxfp4_a_fp8" and self.static_input_scales:
            # W4A8: MXFP4 weights + static FP8 activations
            self.mxfp4_backend, self.experts_cls = select_mxfp4_moe_backend(
                moe, activation_key=kFp8StaticTensorSym
            )
        elif self.ocp_mx_scheme == "w_mxfp4_a_mxfp4":
            # W4A4: MXFP4 weights + MXFP4 activations
            self.mxfp4_backend, self.experts_cls = select_mxfp4_moe_backend(
                moe, activation_key=kMxfp4Dynamic
            )

        # Validation for unsupported schemes
        if any(
            self.ocp_mx_scheme.endswith(a_scheme)
            for a_scheme in ["a_mxfp4", "a_mxfp6_e3m2", "a_mxfp6_e2m3"]
        ):
            if self.static_input_scales:
                raise NotImplementedError(
                    "QuarkOCP_MX_MoEMethod with static input scales is currently "
                    f"not implemented for OCP MX scheme {self.ocp_mx_scheme}. "
                    "Please open an issue."
                )
        elif self.ocp_mx_scheme.endswith("a_fp8") and not self.static_input_scales:
            raise NotImplementedError(
                "QuarkOCP_MX_MoEMethod with dynamic input scales is currently "
                f"not implemented for OCP MX scheme {self.ocp_mx_scheme}. "
                "Please open an issue."
            )

        self.model_type = getattr(
            get_current_vllm_config().model_config.hf_config, "model_type", None
        )

        # If no native backend available, use emulation.
        if self.mxfp4_backend is Mxfp4MoeBackend.NONE:
            self.mxfp4_backend = Mxfp4MoeBackend.EMULATION

        self.experts_cls = backend_to_kernel_cls(self.mxfp4_backend)[0]

        logger.info_once(
            f"Using {self.mxfp4_backend.value} backend for {self.ocp_mx_scheme}"
        )

    def maybe_roundup_sizes(
        self,
        hidden_size: int,
        intermediate_size_per_partition: int,
        act_dtype: torch.dtype,
        moe_parallel_config: FusedMoEParallelConfig,
    ) -> tuple[int, int]:
        hidden_size, intermediate_size_per_partition = super().maybe_roundup_sizes(
            hidden_size=hidden_size,
            intermediate_size_per_partition=intermediate_size_per_partition,
            act_dtype=act_dtype,
            moe_parallel_config=moe_parallel_config,
        )
        # Round per-partition sizes up to each backend's requirement. Emulation is
        # handled inside the helper too (OCP MX block alignment), so no special-case.
        if self.mxfp4_backend is not None:
            hidden_size, intermediate_size_per_partition = (
                mxfp4_round_up_hidden_size_and_intermediate_size(
                    self.mxfp4_backend, hidden_size, intermediate_size_per_partition
                )
            )
        return hidden_size, intermediate_size_per_partition

    def get_packed_dim(self, dim: int, quant_dtype: str):
        if quant_dtype == "mxfp4":
            assert dim % 2 == 0
            return dim // 2
        else:
            # FP6 packs 4 * 6 = 24 bits on 3 bytes.
            assert (dim * 3) % 4 == 0
            return (dim * 3) // 4

    def create_weights(
        self,
        layer: RoutedExperts,
        num_experts: int,
        hidden_size: int,
        intermediate_size_per_partition: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ):
        # Add the quantization method used (per tensor/grouped/channel)
        # to ensure the weight scales are loaded in properly
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.BLOCK.value}
        )

        params_dtype = torch.uint8

        # WEIGHTS
        w13_weight = torch.nn.Parameter(
            torch.zeros(
                num_experts,
                self.moe.w13_num_shards * intermediate_size_per_partition,
                self.get_packed_dim(hidden_size, self.weight_dtype),
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight", w13_weight)

        set_weight_attrs(w13_weight, extra_weight_attrs)

        w2_weight = torch.nn.Parameter(
            torch.zeros(
                num_experts,
                hidden_size,
                self.get_packed_dim(intermediate_size_per_partition, self.weight_dtype),
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_weight", w2_weight)

        set_weight_attrs(w2_weight, extra_weight_attrs)

        # WEIGHT_SCALES
        w13_weight_scale = torch.nn.Parameter(
            torch.ones(
                num_experts,
                self.moe.w13_num_shards * intermediate_size_per_partition,
                hidden_size // OCP_MX_BLOCK_SIZE,
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        w2_weight_scale = torch.nn.Parameter(
            torch.ones(
                num_experts,
                hidden_size,
                intermediate_size_per_partition // OCP_MX_BLOCK_SIZE,
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        set_weight_attrs(w2_weight_scale, extra_weight_attrs)
        set_weight_attrs(w13_weight_scale, extra_weight_attrs)

        layer.register_parameter("w13_weight_scale", w13_weight_scale)
        layer.register_parameter("w2_weight_scale", w2_weight_scale)

        if self.has_bias:
            w13_bias = torch.nn.Parameter(
                torch.zeros(
                    num_experts,
                    self.moe.w13_num_shards * intermediate_size_per_partition,
                    dtype=torch.float32,
                ),
                requires_grad=False,
            )
            layer.register_parameter("w13_bias", w13_bias)
            set_weight_attrs(w13_bias, extra_weight_attrs)

            w2_bias = torch.nn.Parameter(
                torch.zeros(num_experts, hidden_size, dtype=torch.float32),
                requires_grad=False,
            )
            layer.register_parameter("w2_bias", w2_bias)
            set_weight_attrs(w2_bias, extra_weight_attrs)
        else:
            layer.w13_bias, layer.w2_bias = None, None

        # INPUT_SCALES
        if self.static_input_scales:
            w13_input_scale = torch.nn.Parameter(
                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
            )
            layer.register_parameter("w13_input_scale", w13_input_scale)
            set_weight_attrs(w13_input_scale, extra_weight_attrs)

            w2_input_scale = torch.nn.Parameter(
                torch.ones(num_experts, dtype=torch.float32), requires_grad=False
            )
            layer.register_parameter("w2_input_scale", w2_input_scale)
            set_weight_attrs(w2_input_scale, extra_weight_attrs)
        else:
            layer.w13_input_scale = None
            layer.w2_input_scale = None

    def process_weights_after_loading(self, layer):
        self._setup_kernel(layer)

    def _setup_kernel(self, layer: RoutedExperts):
        """Setup kernel using oracle functions for MXFP4 schemes (W4A16, W4A8)."""
        w13_bias = getattr(layer, "w13_bias", None)
        w2_bias = getattr(layer, "w2_bias", None)

        # Convert weights to kernel format (handles all backend-specific logic)
        w13, w2, w13_scale, w2_scale, w13_bias, w2_bias = (
            convert_gpt_oss_weight_to_mxfp4_moe_kernel_format(
                mxfp4_backend=self.mxfp4_backend,
                layer=layer,
                w13_weight=layer.w13_weight,
                w2_weight=layer.w2_weight,
                w13_weight_scale=layer.w13_weight_scale,
                w2_weight_scale=layer.w2_weight_scale,
                w13_bias=w13_bias,
                w2_bias=w2_bias,
                w13_input_scale=layer.w13_input_scale,
                w2_input_scale=layer.w2_input_scale,
            )
        )

        # Handle weight/scale assignment based on backend type
        if self.mxfp4_backend in TRITON_BACKENDS or self.mxfp4_backend in (
            Mxfp4MoeBackend.AITER_MXFP4_FP8,
        ):
            # Triton-based backends: w13/w2 are triton_kernels.tensor.Tensor
            # Store on layer for apply(), scales are PrecisionConfig
            layer.w13_weight = w13
            layer.w2_weight = w2
            self.w13_precision_config = w13_scale
            self.w2_precision_config = w2_scale
        else:
            # Standard backends: replace parameters
            replace_parameter(layer, "w13_weight", w13)
            replace_parameter(layer, "w2_weight", w2)
            replace_parameter(layer, "w13_weight_scale", w13_scale)
            replace_parameter(layer, "w2_weight_scale", w2_scale)

        if w13_bias is not None and w2_bias is not None:
            replace_parameter(layer, "w13_bias", w13_bias)
            replace_parameter(layer, "w2_bias", w2_bias)

        torch.accelerator.empty_cache()

        # Build quant config and kernel
        self.moe_quant_config = self.get_fused_moe_quant_config(layer)
        if self.moe_quant_config is not None and self.experts_cls is not None:
            self.moe_kernel = make_mxfp4_moe_kernel(
                moe_quant_config=self.moe_quant_config,
                moe_config=self.moe,
                mxfp4_backend=self.mxfp4_backend,
                experts_cls=self.experts_cls,
                routing_tables=layer._expert_routing_tables(),
            )
            self.moe_kernel.fused_experts.process_weights_after_loading(layer)

    def get_fused_moe_quant_config(
        self, layer: RoutedExperts
    ) -> FusedMoEQuantConfig | None:
        # For oracle-based backends (W4A16, W4A8), use make_mxfp4_moe_quant_config
        if self.mxfp4_backend not in (Mxfp4MoeBackend.NONE, Mxfp4MoeBackend.EMULATION):
            # Determine scale source based on backend type
            if self.mxfp4_backend in TRITON_BACKENDS or self.mxfp4_backend in (
                Mxfp4MoeBackend.AITER_MXFP4_FP8,
            ):
                w1_scale = self.w13_precision_config
                w2_scale = self.w2_precision_config
            else:
                w1_scale = layer.w13_weight_scale
                w2_scale = layer.w2_weight_scale

            return make_mxfp4_moe_quant_config(
                mxfp4_backend=self.mxfp4_backend,
                w1_scale=w1_scale,
                w2_scale=w2_scale,
                w1_bias=getattr(layer, "w13_bias", None),
                w2_bias=getattr(layer, "w2_bias", None),
                a1_scale=getattr(layer, "w13_input_scale", None),
                a2_scale=getattr(layer, "w2_input_scale", None),
                gemm1_alpha=getattr(layer, "swiglu_alpha", None),
                gemm1_beta=getattr(layer, "swiglu_beta", None),
                swiglu_limit=getattr(layer, "swiglu_limit", None),
                layer=layer,
            )

        # Emulation and other schemes
        if self.ocp_mx_scheme == "w_mxfp4":
            return mxfp4_w4a16_moe_quant_config(
                w1_scale=layer.w13_weight_scale,
                w2_scale=layer.w2_weight_scale,
                w1_bias=layer.w13_bias,
                w2_bias=layer.w2_bias,
            )
        elif self.ocp_mx_scheme == "w_mxfp4_a_fp8":
            return mxfp4_w4a8_moe_quant_config(
                w1_scale=layer.w13_weight_scale,
                w2_scale=layer.w2_weight_scale,
                a1_scale=layer.w13_input_scale,
                a2_scale=layer.w2_input_scale,
                w1_bias=layer.w13_bias,
                w2_bias=layer.w2_bias,
                block_shape=None,
            )
        elif self.ocp_mx_scheme in ["w_mxfp6_e3m2_a_fp8", "w_mxfp6_e2m3_a_fp8"]:
            raise NotImplementedError(
                "Currently there is no corresponding fused moe quant config configured "
                f"in vLLM for OCP MX scheme {self.ocp_mx_scheme}. Please open an issue."
            )
        else:
            return ocp_mx_moe_quant_config(
                quant_dtype=self.input_dtype,
                weight_dtype=self.weight_dtype,
                w1_scale=layer.w13_weight_scale,
                w2_scale=layer.w2_weight_scale,
                w1_bias=layer.w13_bias,
                w2_bias=layer.w2_bias,
                a1_scale=None,
                a2_scale=None,
                block_shape=None,
                gemm1_alpha=getattr(layer, "swiglu_alpha", None),
                gemm1_beta=getattr(layer, "swiglu_beta", None),
                gemm1_clamp_limit=getattr(layer, "swiglu_limit", None),
            )

    @property
    def supports_eplb(self) -> bool:
        # AITER shuffle keeps expert dim outermost, so EPLB row moves are layout-safe.
        return True

    @property
    def is_monolithic(self) -> bool:
        if self.moe_kernel is not None:
            return self.moe_kernel.is_monolithic
        return False

    def apply(
        self,
        layer: RoutedExperts,
        x: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        shared_experts: SharedExperts | None,
        shared_experts_input: torch.Tensor | None,
    ) -> torch.Tensor:
        assert self.moe_kernel is not None
        return self.moe_kernel.apply(
            hidden_states=x,
            w1=layer.w13_weight,
            w2=layer.w2_weight,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            activation=layer.activation,
            global_num_experts=layer.global_num_experts,
            apply_router_weight_on_input=layer.apply_router_weight_on_input,
            expert_map=layer.expert_map,
            shared_experts_input=shared_experts_input,
        )

    def apply_monolithic(
        self,
        layer: RoutedExperts,
        x: torch.Tensor,
        router_logits: torch.Tensor,
        input_ids: torch.Tensor | None = None,
    ) -> torch.Tensor:
        assert self.is_monolithic
        assert self.moe_kernel is not None
        return self.moe_kernel.apply_monolithic(
            hidden_states=x,
            w1=layer.w13_weight,
            w2=layer.w2_weight,
            router_logits=router_logits,
            activation=layer.activation,
            global_num_experts=layer.global_num_experts,
            expert_map=layer.expert_map,
            apply_router_weight_on_input=layer.apply_router_weight_on_input,
        )

class QuarkW8A8Int8MoEMethod(QuarkMoEMethod):
    """Quark W8A8 INT8 MoE method."""

    def __init__(
        self,
        weight_config: dict[str, Any],
        input_config: dict[str, Any],
        moe: FusedMoEConfig,
    ):
        super().__init__(moe)
        self.weight_quant = weight_config
        self.input_quant = input_config
        self.weight_qscheme = self.weight_quant.get("qscheme", "per_tensor")
        self.static_input_scales = not self.input_quant.get("is_dynamic", False)

        self.moe_quant_config: FusedMoEQuantConfig | None = None
        self.moe_kernel: mk.FusedMoEKernel | None = None
        self.int8_backend: Int8MoeBackend | None = None
        self.experts_cls: type[mk.FusedMoEExperts] | None = None

        # Dynamic-activation INT8 MoE goes through the oracle + modular kernel.
        # The modular TritonExperts kernel consumes float activations and
        # quantizes them to int8 itself, so it cannot apply a loaded static
        # activation scale (this matches CompressedTensorsW8A8Int8MoEMethod).
        # TODO: Static-activation INT8 therefore stays on the legacy fused_experts
        # path (see apply()) for now, preserving pre-refactor behavior.
        # Needs to be migrated to expert backend.
        if not self.static_input_scales:
            # Map the Quark weight scheme to oracle quant keys. Per-channel
            # weights pair with dynamic per-token activations; per-tensor
            # weights with dynamic per-tensor activations.
            if self.weight_qscheme == "per_channel":
                weight_key = kInt8StaticChannelSym
                activation_key = kInt8DynamicTokenSym
            else:
                weight_key = kInt8StaticTensorSym
                activation_key = kInt8DynamicTensorSym

            self.int8_backend, self.experts_cls = select_int8_moe_backend(
                config=moe,
                weight_key=weight_key,
                activation_key=activation_key,
            )

    def create_weights(
        self,
        layer: torch.nn.Module,
        num_experts: int,
        hidden_size: int,
        intermediate_size_per_partition: int,
        params_dtype: torch.dtype,
        **extra_weight_attrs,
    ):
        layer.num_experts = num_experts
        layer.orig_dtype = params_dtype
        layer.weight_block_size = None
        params_dtype = torch.int8

        # WEIGHTS
        w13_weight = torch.nn.Parameter(
            torch.empty(
                num_experts,
                self.moe.w13_num_shards * intermediate_size_per_partition,
                hidden_size,
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight", w13_weight)
        set_weight_attrs(w13_weight, extra_weight_attrs)

        w2_weight = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                intermediate_size_per_partition,
                dtype=params_dtype,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_weight", w2_weight)
        set_weight_attrs(w2_weight, extra_weight_attrs)

        # WEIGHT_SCALES
        if self.weight_qscheme == "per_channel":
            w13_weight_scale = torch.nn.Parameter(
                torch.ones(
                    num_experts,
                    self.moe.w13_num_shards * intermediate_size_per_partition,
                    dtype=torch.float32,
                ),
                requires_grad=False,
            )
            layer.register_parameter("w13_weight_scale", w13_weight_scale)
            w2_weight_scale = torch.nn.Parameter(
                torch.ones(num_experts, hidden_size, dtype=torch.float32),
                requires_grad=False,
            )
            layer.register_parameter("w2_weight_scale", w2_weight_scale)
            extra_weight_attrs.update(
                {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value}
            )
            set_weight_attrs(w13_weight_scale, extra_weight_attrs)
            set_weight_attrs(w2_weight_scale, extra_weight_attrs)
        else:
            # per-tensor: one scalar per expert (two for the fused w1/w3)
            w13_weight_scale = torch.nn.Parameter(
                torch.ones(num_experts, self.moe.w13_num_shards, dtype=torch.float32),
                requires_grad=False,
            )
            layer.register_parameter("w13_weight_scale", w13_weight_scale)
            w2_weight_scale = torch.nn.Parameter(
                torch.ones(num_experts, dtype=torch.float32),
                requires_grad=False,
            )
            layer.register_parameter("w2_weight_scale", w2_weight_scale)
            extra_weight_attrs.update(
                {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
            )
            set_weight_attrs(w13_weight_scale, extra_weight_attrs)
            set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        # INPUT_SCALES
        if self.static_input_scales:
            # Static activations: the per-expert scales are loaded from the
            # checkpoint (used by the legacy fused_experts path).
            w13_input_scale = torch.nn.Parameter(
                torch.ones(num_experts, dtype=torch.float32),
                requires_grad=False,
            )
            layer.register_parameter("w13_input_scale", w13_input_scale)
            set_weight_attrs(w13_input_scale, extra_weight_attrs)

            w2_input_scale = torch.nn.Parameter(
                torch.ones(num_experts, dtype=torch.float32),
                requires_grad=False,
            )
            layer.register_parameter("w2_input_scale", w2_input_scale)
            set_weight_attrs(w2_input_scale, extra_weight_attrs)
        else:
            # Dynamic activations are quantized in-kernel (no stored scale).
            layer.w13_input_scale = None
            layer.w2_input_scale = None

        # ZERO POINTS (loaded but discarded after loading; kernel uses symmetric)
        w13_input_zero_point = torch.nn.Parameter(
            torch.zeros(num_experts, self.moe.w13_num_shards, dtype=torch.int8),
            requires_grad=False,
        )
        layer.register_parameter("w13_input_zero_point", w13_input_zero_point)
        set_weight_attrs(w13_input_zero_point, extra_weight_attrs)

        w2_input_zero_point = torch.nn.Parameter(
            torch.zeros(num_experts, dtype=torch.int8),
            requires_grad=False,
        )
        layer.register_parameter("w2_input_zero_point", w2_input_zero_point)
        set_weight_attrs(w2_input_zero_point, extra_weight_attrs)

        if self.weight_qscheme == "per_channel":
            w13_weight_zero_point = torch.nn.Parameter(
                torch.zeros(
                    num_experts,
                    self.moe.w13_num_shards * intermediate_size_per_partition,
                    dtype=torch.int8,
                ),
                requires_grad=False,
            )
            w2_weight_zero_point = torch.nn.Parameter(
                torch.zeros(num_experts, hidden_size, dtype=torch.int8),
                requires_grad=False,
            )
        else:
            w13_weight_zero_point = torch.nn.Parameter(
                torch.zeros(num_experts, self.moe.w13_num_shards, dtype=torch.int8),
                requires_grad=False,
            )
            w2_weight_zero_point = torch.nn.Parameter(
                torch.zeros(num_experts, dtype=torch.int8),
                requires_grad=False,
            )
        layer.register_parameter("w13_weight_zero_point", w13_weight_zero_point)
        set_weight_attrs(w13_weight_zero_point, extra_weight_attrs)
        layer.register_parameter("w2_weight_zero_point", w2_weight_zero_point)
        set_weight_attrs(w2_weight_zero_point, extra_weight_attrs)

        # BIAS
        if self.has_bias:
            w13_bias = torch.nn.Parameter(
                torch.zeros(
                    num_experts,
                    self.moe.w13_num_shards * intermediate_size_per_partition,
                    dtype=torch.float32,
                ),
                requires_grad=False,
            )
            layer.register_parameter("w13_bias", w13_bias)
            set_weight_attrs(w13_bias, extra_weight_attrs)
            w2_bias = torch.nn.Parameter(
                torch.zeros(num_experts, hidden_size, dtype=torch.float32),
                requires_grad=False,
            )
            layer.register_parameter("w2_bias", w2_bias)
            set_weight_attrs(w2_bias, extra_weight_attrs)
        else:
            layer.w13_bias, layer.w2_bias = None, None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        # Discard zero points (INT8 fused MoE kernel uses symmetric quant)
        for attr in (
            "w13_input_zero_point",
            "w2_input_zero_point",
            "w13_weight_zero_point",
            "w2_weight_zero_point",
        ):
            if hasattr(layer, attr):
                delattr(layer, attr)

        # For static input scales, collapse the per-expert scales to a single
        # value (the legacy fused_experts path expects one scale per layer).
        if self.static_input_scales:
            if layer.w13_input_scale is None or layer.w2_input_scale is None:
                raise ValueError(
                    "QuantConfig has static quantization, but found "
                    "activation scales are None."
                )
            if not all_close_1d(layer.w13_input_scale) or not all_close_1d(
                layer.w2_input_scale
            ):
                logger.warning_once(
                    "Found input_scales that are not equal for "
                    "INT8 MoE layer. Using the maximum across experts "
                    "for each layer."
                )
            layer.w13_input_scale = torch.nn.Parameter(
                layer.w13_input_scale.max(), requires_grad=False
            )
            layer.w2_input_scale = torch.nn.Parameter(
                layer.w2_input_scale.max(), requires_grad=False
            )

        # Per-channel scales: 2D [E, N] -> 3D [E, N, 1] for the int8 MoE kernel.
        if self.weight_qscheme == "per_channel":
            for attr in ("w13_weight_scale", "w2_weight_scale"):
                param = getattr(layer, attr, None)
                if param is not None and param.dim() == 2:
                    replace_parameter(
                        layer,
                        attr,
                        torch.nn.Parameter(
                            param.data.unsqueeze(-1).contiguous(),
                            requires_grad=False,
                        ),
                    )

        # For per-tensor weights, merge the w1/w3 scales into a single
        # per-expert scale (dequant -> requant at the max scale).
        if self.weight_qscheme == "per_tensor":
            assert layer.w13_weight_scale is not None
            shard_size = layer.intermediate_size_per_partition
            max_w13_scales = layer.w13_weight_scale.max(dim=1).values

            for expert_id in range(layer.local_num_experts):
                start = 0
                for shard_id in range(self.moe.w13_num_shards):
                    dq_weight = per_tensor_dequantize(
                        layer.w13_weight[expert_id][start : start + shard_size, :],
                        layer.w13_weight_scale[expert_id][shard_id],
                    )
                    layer.w13_weight[expert_id][start : start + shard_size, :], _, _ = (
                        ops.scaled_int8_quant(
                            dq_weight,
                            scale=max_w13_scales[expert_id],
                        )
                    )
                    start += shard_size

            layer.w13_weight_scale = torch.nn.Parameter(
                max_w13_scales, requires_grad=False
            )

        # Dynamic activations run through the oracle's modular kernel; static
        # activations use the legacy fused_experts path in apply().
        if not self.static_input_scales:
            assert self.int8_backend is not None
            assert self.experts_cls is not None
            w13, w2 = convert_to_int8_moe_kernel_format(
                int8_backend=self.int8_backend,
                w13=layer.w13_weight,
                w2=layer.w2_weight,
                layer=layer,
                w13_scale=layer.w13_weight_scale,
            )
            replace_parameter(layer, "w13_weight", w13)
            replace_parameter(layer, "w2_weight", w2)

        self.moe_quant_config = self.get_fused_moe_quant_config(layer)
        assert self.moe_quant_config is not None

        if not self.static_input_scales:
            assert self.int8_backend is not None
            assert self.experts_cls is not None
            self.moe_kernel = make_int8_moe_kernel(
                int8_backend=self.int8_backend,
                moe_quant_config=self.moe_quant_config,
                moe_config=self.moe,
                experts_cls=self.experts_cls,
                routing_tables=layer._expert_routing_tables(),
            )

    def get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        # Static-activation INT8 has no oracle backend (it uses the legacy
        # fused_experts path); build its config directly.
        if self.int8_backend is None:
            return int8_w8a8_moe_quant_config(
                w1_scale=layer.w13_weight_scale,
                w2_scale=layer.w2_weight_scale,
                a1_scale=layer.w13_input_scale,
                a2_scale=layer.w2_input_scale,
                w1_bias=getattr(layer, "w13_bias", None),
                w2_bias=getattr(layer, "w2_bias", None),
                per_act_token_quant=False,
            )
        return make_int8_moe_quant_config(
            int8_backend=self.int8_backend,
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            a1_scale=layer.w13_input_scale,
            a2_scale=layer.w2_input_scale,
            w1_bias=getattr(layer, "w13_bias", None),
            w2_bias=getattr(layer, "w2_bias", None),
            per_act_token_quant=(self.weight_qscheme == "per_channel"),
            layer=layer,
        )

    def apply(
        self,
        layer: RoutedExperts,
        x: torch.Tensor,
        topk_weights: torch.Tensor,
        topk_ids: torch.Tensor,
        shared_experts: SharedExperts | None,
        shared_experts_input: torch.Tensor | None,
    ) -> torch.Tensor:
        if self.moe_kernel is not None:
            return self.moe_kernel.apply(
                hidden_states=x,
                w1=layer.w13_weight,
                w2=layer.w2_weight,
                topk_weights=topk_weights,
                topk_ids=topk_ids,
                activation=layer.activation,
                global_num_experts=layer.global_num_experts,
                apply_router_weight_on_input=layer.apply_router_weight_on_input,
                expert_map=layer.expert_map,
                shared_experts_input=shared_experts_input,
            )

        # Static-activation INT8 MoE: legacy monolithic path (the modular kernel
        # quantizes activations dynamically and cannot apply a loaded scale).
        from vllm.model_executor.layers.fused_moe import fused_experts

        return fused_experts(
            hidden_states=x,
            w1=layer.w13_weight,
            w2=layer.w2_weight,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            activation=layer.activation,
            apply_router_weight_on_input=layer.apply_router_weight_on_input,
            global_num_experts=layer.global_num_experts,
            expert_map=layer.expert_map,
            quant_config=self.moe_quant_config,
        )